A method and system for predicting urban road network traffic state and congestion propagation probability

Abstract

The present application belongs to the technical field of intelligent transportation, and specifically discloses a kind of urban road network traffic state and congestion propagation probability prediction method and system.The present application first introduces spatial distance, historical traffic correlation and other information on the basis of road topology, forms a weighted checkpoint road network;Secondly, a conditional denoising diffusion model based on checkpoint space-time graph is proposed, which models the conditional distribution of future multi-step graph signals by forward noise addition and reverse denoising process, with historical multi-step graph signals and road network structure as conditions.The present application obtains multiple future traffic evolution sample trajectories by multiple sampling, maps node operation indicators to discrete congestion levels, counts the occurrence frequency of each node at different congestion levels, and obtains node congestion level probability;Further, the time sequence co-occurrence relationship of congestion state between adjacent nodes is counted, the propagation conditional probability of congestion in the road network is estimated, and the key node influence degree index and high-risk propagation path are constructed.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation and relates to a method and system for predicting urban road network traffic conditions and the probability of congestion propagation. Background Technology

[0002] Currently, existing research and engineering practices regarding the utilization of checkpoint data mainly focus on the following areas: short-term traffic flow prediction, road network-level spatiotemporal prediction, congestion propagation and mechanism analysis, and probability prediction and diffusion models.

[0003] In the area of ​​road network-level spatiotemporal prediction, a deterministic prediction scheme based on checkpoint road networks (GCN-GRU type) and a diffusion model prediction scheme based on spatiotemporal maps have been proposed. However, traditional schemes still have the following shortcomings:

[0004] I. Lack of unified and detailed spatiotemporal diagram modeling for checkpoints.

[0005] Traditional prediction schemes only construct simple 0 / 1 adjacency matrices based on road connectivity, resulting in an insufficiently refined characterization of real traffic connections by the graph structure. This directly affects the accuracy of subsequent spatial feature extraction and prediction.

[0006] II. The prediction results are mainly deterministic point estimates, lacking descriptions of uncertainties.

[0007] Mainstream models such as GCN-GRU and LSTM output a single predicted value, and at most provide a rough confidence interval through multi-model ensemble or empirical methods. For a highly uncertain system like urban traffic, a single deterministic point prediction cannot calculate the specific probability of future congestion, nor can it derive the mathematical boundary of the prediction interval.

[0008] III. Disconnection between congestion propagation analysis and prediction models.

[0009] Existing studies on congestion propagation based on epidemiological models or topological analysis often construct propagation equations under simplified assumptions, resulting in weak coupling with actual checkpoint observation sequences. The separate implementation of prediction models and propagation analysis makes it difficult to simultaneously output node states and propagation risks within the same spatiotemporal prediction framework, and lacks a data-driven quantitative characterization of the propagation mechanism.

[0010] IV. Lack of node-level congestion probability and key node impact indicators.

[0011] Existing solutions often focus on the average error of the entire network, the error of a single node, or several aggregated indicators. They rarely output the probability distribution of each checkpoint under different congestion levels, and they also lack methods for calculating the impact of key nodes based on multiple future sample trajectories. It is difficult to obtain information on checkpoints with a high probability of congestion in the future and the congestion propagation path from the model output.

[0012] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0013] The purpose of this invention is to propose a method for predicting urban road network traffic conditions and congestion propagation probability. This method first constructs a spatiotemporal map of urban road network checkpoints based on checkpoint data, and then proposes a conditional denoising diffusion model based on the checkpoint spatiotemporal map. The model can output the probability distribution and prediction interval of node-level traffic conditions while outputting the predicted mean. Finally, under a unified framework, it is conducive to the joint evaluation of node congestion probability and congestion propagation probability.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] A method for predicting urban road network traffic conditions and congestion propagation probability includes the following steps:

[0016] Step 1. Obtain checkpoint data. Preprocess the data to obtain complete and continuous checkpoint operation time series data. Obtain a three-dimensional data tensor through time aggregation and indicator calculation. Simultaneously, construct a weighted adjacency matrix. ;

[0017] Step 2. Construct a conditional denoising diffusion prediction model and utilize three-dimensional data tensors and The training samples are constructed, the model is trained, and then the trained model is used to predict future traffic evolution trajectories.

[0018] First, a set of pure noise data with the same time and spatial dimensions as the prediction target is randomly selected from the standard Gaussian space;

[0019] Then, this set of pure noise data and the latest historical feature sequence... , When the number of steps and the embedding vector n are input together into the trained model, the model outputs a future traffic evolution trajectory that conforms to physical laws.

[0020] Maintain the latest historical feature sequence and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of the following is a unique future transportation evolution trajectory, i.e., a sample trajectory.

[0021] Step 3. Map the index of each node and each time step in the sample trajectory to a discrete congestion level, and count the frequency of occurrence of each level to obtain the probability of node congestion level and the corresponding prediction interval.

[0022] Step 4. Based on multiple future sample trajectories, statistically analyze the co-occurrence relationship of congestion states at adjacent checkpoints under different time lags, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the impact degree of node congestion propagation through path probability accumulation statistics to identify key nodes of congestion propagation.

[0023] Furthermore, based on the aforementioned method for predicting urban road network traffic conditions and congestion propagation probability, this invention also proposes a corresponding urban road network traffic conditions and congestion propagation probability prediction system, which adopts the following technical solution:

[0024] A system for predicting urban road network traffic conditions and congestion propagation probability includes the following modules:

[0025] The preprocessing module is used to acquire checkpoint data. Through data preprocessing, it obtains complete and continuous checkpoint operation time series data, and obtains a three-dimensional data tensor through time aggregation and index calculation; at the same time, it constructs a weighted adjacency matrix. ;

[0026] The trajectory prediction module is used to construct a conditional denoising diffusion prediction model, utilizing three-dimensional data tensors and... The model is trained using the constructed training samples; then the trained model is used to predict future traffic evolution trajectories.

[0027] First, a set of pure noise data with the same time and spatial dimensions as the prediction target is randomly selected from the standard Gaussian space;

[0028] This set of pure noise data and the latest historical feature sequences , When the number of steps and the embedding vector n are input together into the trained model, the model outputs a future traffic evolution trajectory that conforms to physical laws.

[0029] Then keep the latest historical feature sequence. and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of the following is a unique future transportation evolution trajectory, i.e., a sample trajectory.

[0030] The urban road network traffic condition prediction module is used to map the index of each node and each time step in the sample trajectory to discrete congestion levels, and to count the frequency of occurrence of each level, thereby obtaining the probability of node congestion level and the corresponding prediction interval.

[0031] The module also includes a congestion propagation probability prediction module, which is used to statistically analyze the co-occurrence relationship of congestion states at different time lags of adjacent checkpoints based on multiple future sample trajectories, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the impact degree of node congestion propagation through path probability accumulation statistics, and identify key nodes of congestion propagation.

[0032] The present invention has the following advantages:

[0033] As described above, this invention discloses a method for predicting urban road network traffic conditions and congestion propagation probability. This method represents the urban checkpoint road network as a checkpoint spatiotemporal map containing spatial topological relationships and dynamic features. A conditional denoising diffusion prediction model is constructed on this checkpoint spatiotemporal map. Historical multi-step map signals and weighted adjacency matrices are used as conditional inputs to the model, while future multi-step map signals are considered as targets to be generated from noise. Forward denoising and reverse denoising processes are introduced to directly model the "conditional probability distribution of future traffic conditions" on the checkpoint spatiotemporal map, rather than using traditional single-point prediction. In the model denoising stage, a U-shaped denoising network ST-GUNet structure combining graph convolution and temporal modeling is used. To adapt to checkpoint spatiotemporal map signals, a graph-structured U-shaped denoising network is designed in the diffusion model: on the one hand, a graph convolution module is used to extract spatial correlations using a weighted adjacency matrix; on the other hand, temporal modules such as GRU and TCN are introduced to characterize historical evolution features. Furthermore, U-Net-style multi-scale encoding-decoding and skip connections are used to fuse spatiotemporal features at different scales for denoising at each diffusion step. Furthermore, a method for constructing a probability distribution of node congestion levels based on multiple diffusion sampling is proposed. This method enables probabilistic prediction of urban road network traffic conditions by constructing a probability distribution of node congestion levels through multiple diffusion sampling. Under the same historical conditions and road network structure, multiple sample trajectories of future traffic conditions are generated through multiple diffusion sampling. The indicators of each node and each time step in the samples are mapped to discrete congestion levels (such as smooth traffic, light congestion, severe congestion, etc.), and the frequency of occurrence of each level is statistically analyzed to obtain the probability of node congestion levels and the corresponding prediction intervals. Further, this invention also provides a method for calculating the probability of congestion propagation and the influence of key nodes based on sample trajectories. By statistically analyzing the co-occurrence relationship of congestion states at adjacent checkpoints under different time lags based on multiple future sample trajectories, the conditional probability of congestion propagating from upstream nodes to downstream nodes is estimated. Through cumulative path probability statistics, the influence of node congestion propagation is defined to identify key nodes and high-risk paths for congestion propagation. Attached Figure Description

[0034] Figure 1 This is a flowchart of the method for predicting urban road network traffic conditions and congestion propagation probability in an embodiment of the present invention;

[0035] Figure 2This is a schematic diagram of the construction of the checkpoint timing tensor in an embodiment of the present invention;

[0036] Figure 3 This is a schematic diagram illustrating the construction of the weighted adjacency matrix W in an embodiment of the present invention;

[0037] Figure 4 This is a diagram illustrating the training principle of the conditional diffusion model in this embodiment of the invention.

[0038] Figure 5 This is a schematic diagram of the denoising network structure in an embodiment of the present invention;

[0039] Figure 6 This is a network structure diagram of the spatiotemporal residual block in an embodiment of the present invention;

[0040] Figure 7 This is a probability output diagram of inference and multiple sampling in an embodiment of the present invention;

[0041] Figure 8 This is a flowchart illustrating the congestion propagation simulation and key node identification in an embodiment of the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0043] Example 1

[0044] This embodiment 1 describes a method for predicting urban road network traffic conditions and congestion propagation probability. First, it proposes a method for constructing a spatiotemporal map of the urban road network based on checkpoint data. On the basis of road topology, it introduces information such as spatial distance and historical traffic correlation to form a weighted checkpoint road network, enabling the map structure to more realistically reflect the actual traffic connections between checkpoints and providing a more reasonable structural prior for subsequent probability prediction and propagation analysis. Second, this invention proposes a conditional denoising diffusion model based on the road network spatiotemporal map. On the checkpoint spatiotemporal map, using historical multi-step map signals and road network structure as conditions, it models the conditional distribution of future multi-step map signals through forward noise addition and reverse denoising processes. Multiple diffusion sampling is performed under the same historical conditions, utilizing the randomness of the initial noise to generate multiple traffic sample trajectories reflecting different evolution trends. Furthermore, this invention can simultaneously achieve joint evaluation of node congestion probability and congestion propagation probability within a unified framework. First, by obtaining multiple sample trajectories of future traffic evolution through repeated sampling, node operation indicators are mapped to discrete congestion levels. The frequency of occurrence of each node under different congestion levels is statistically analyzed to obtain the probability of node congestion level. Furthermore, the temporal co-occurrence relationship of congestion states between adjacent nodes is statistically analyzed to estimate the conditional probability of congestion propagation in the road network, constructing key node influence indicators and high-risk propagation paths. This invention provides an engineering-featured congestion propagation probability prediction scheme for refined management of urban road networks without altering existing checkpoint deployments.

[0045] like Figure 1 As shown, the method for predicting urban road network traffic conditions and congestion propagation probability in this embodiment includes the following steps:

[0046] Step 1. Obtain checkpoint data. Preprocess the data to obtain complete and continuous checkpoint operation time series data. Then, obtain a three-dimensional tensor dataset through time aggregation and index calculation, such as... Figure 2 As shown. A weighted adjacency matrix is ​​also constructed. .

[0047] Step 1.1. Checkpoint data collection.

[0048] Vehicle passage records are obtained in real time from each checkpoint in the high-definition checkpoint system.

[0049] Each original record should include: a unique checkpoint number, the precise time of vehicle capture (milliseconds), the de-identified license plate number, the vehicle type code (e.g., large vehicle / small vehicle), the lane number, and the direction of travel.

[0050] All original records are stored in a distributed database in timestamp order, serving as the underlying data source for subsequent analysis.

[0051] Step 1.2. Data cleaning and missing data handling.

[0052] The time series data is cleaned and completed, and obvious errors and outliers are deleted, such as inverted timestamps and empty license plate fields. Abnormally large traffic spikes and abnormally low values ​​are identified using threshold or statistical detection methods. For individual missing time windows, interpolation of nearby times and weighted estimation of adjacent checkpoints are used to complete the data.

[0053] After the above preprocessing, relatively complete and continuous checkpoint operation time series data are obtained.

[0054] Step 1.3. Time aggregation and indicator calculation.

[0055] Set a fixed time interval , For example, set it to 5 minutes.

[0056] For each checkpoint and each time window, the vehicle passage records are statistically analyzed, and the following indicators are calculated: traffic flow within the time window (the total number of all vehicles passing through the checkpoint within the time window) and average speed within the time window.

[0057] Specifically, by using "license plate matching" technology, the system can identify vehicles that have passed through adjacent checkpoints in succession. and Capture timestamp and Then the speed of the vehicle between two adjacent checkpoints for:

[0058] ,in This is the shortest physical distance between the two checkpoints calculated for subsequent steps. All successfully matched vehicle samples within a 5-minute period are analyzed, and their harmonic average speed is calculated to obtain the average speed for that road segment during that time period.

[0059] Target vehicles are located using license plate matching technology within each fixed time window (e.g., 5 minutes). These vehicles must pass through the starting checkpoint and the adjacent downstream checkpoint sequentially within this time window.

[0060] Extract the specific timestamps of the same vehicle captured at two checkpoints, calculate the vehicle's speed over the interval by combining the actual physical distance between the two checkpoints, and assign this average speed directly to the starting checkpoint.

[0061] The indicator dimension F includes two core operational indicators, preferably traffic flow and average vehicle speed. This forms a three-dimensional data tensor of "time × checkpoint × indicator". This serves as a basic description of the road network's operational status.

[0062] First dimension This represents a historical time series divided into 5-minute steps; the second dimension... Represents the set of all observation checkpoints within the region; third dimension It represents two core operational indicators: traffic flow and average speed.

[0063] Step 1.4. Construct a weighted spatiotemporal representation of the road network.

[0064] This invention constructs a directed weighted graph, or weighted adjacency matrix, that can realistically reflect the operational characteristics of urban road networks by combining "physical connection-based structure" and "multi-source attribute-based weighting". .

[0065] like Figure 3 As shown, The specific construction process is divided into two stages: initial topology establishment and comprehensive weight calculation.

[0066] I. Establish the initial topology.

[0067] Each checkpoint is considered a node in a graph, and all checkpoints constitute a set of nodes. .

[0068] in Total number of checkpoints; , … Each node represents a node.

[0069] Based on urban road centerline data and traffic flow information, if vehicles exit from the checkpoint... After exiting, it can reach the checkpoint directly without going through other checkpoints. Then it is determined that there exists a directed edge. .

[0070] The above judgment rules indicate that, i.e., checkpoint and checkpoint Located upstream or downstream of the same or adjacent road segment.

[0071] A basic topological adjacency matrix is ​​constructed based on the above determination rules. .

[0072] in , indicating checkpoint With checkpoint There is a connection. This indicates that there is no direct connection;

[0073] II. Calculation of overall weight.

[0074] In the basic topological adjacency matrix Based on this, spatial distance is introduced. Traffic correlation and road function correction coefficient Construct a weighted adjacency matrix ,in medium elements The formula is expressed as follows:

[0075] .

[0076] Where α is a hyperparameter, and its value ranges from [0, 1]. The bandwidth threshold for the Gaussian kernel function; For topology indicator functions, only Weights are calculated between nodes that are physically connected; otherwise, they are set to 0.

[0077] The following details spatial distance. Traffic correlation and road function correction coefficient The process of obtaining etc.

[0078] II.1. Spatial Distance calculate.

[0079] The Dijkstra algorithm is used to find the shortest path between two points in a road network topology, as shown in the following formula:

[0080] .

[0081] in, Indicates from the checkpoint to checkpoint The set of all feasible physical paths; For a given path in the set, it consists of a series of consecutive road segments. composition; Indicates road segment The actual physical length (unit: meters).

[0082] Specifically, if the checkpoint and If the two are topologically inaccessible (e.g., belonging to different enclosed areas or restricted by one-way streets), then set... =∞, corresponding to the subsequent weight calculation. The value is assigned to 0.

[0083] II.2. Historical Flow Correlation calculate.

[0084] Extract the checkpoints separately and In historical statistical periods Traffic flow time series within the area, constructing observation vectors and .

[0085] in and Refers to the checkpoint and at any time step Specific flow observation values.

[0086] The Pearson correlation coefficient is used to calculate the degree of linear correlation between the two in terms of temporal fluctuations, and to identify the associated node pairs that may be spatially far apart but whose congestion time and flow change patterns are highly consistent (such as key bottlenecks in the upstream and downstream).

[0087] The calculation formula is as follows:

[0088] .

[0089] in and These are the arithmetic averages of the flow rates of the two traffic streams within the statistical period T.

[0090] The calculated correlation coefficient The value range is [-1, 1], and its physical meaning lies in quantifying the degree of synchronization between nodes.

[0091] The closer the value is to 1, the stronger the positive correlation and synchronicity of the traffic flow status of the two checkpoints (such as the outbreak of congestion during peak hours or the smooth dissipation during off-peak hours); if the value is close to 0, it means that the operation patterns of the two do not interfere with each other.

[0092] II.3. Construction of the comprehensive edge weight matrix.

[0093] After quantifying the spatial geometric proximity between nodes (i.e. ), and the correlation of traffic evolution (i.e. ) and road functional attributes (i.e. Based on this, a multi-view fusion strategy is adopted.

[0094] This invention includes a "geographic view" (composed of spatial distance). The decision reflects physical proximity) and the "data view" (based on traffic correlation). (The decision reflects the synchronicity of the operational rules).

[0095] By fusing these two dimensions, the matrix not only represents how the roads are connected, but also knows which nodes are functionally coupled, thus constructing the final weighted adjacency matrix W.

[0096] To balance the dual characteristics of "geographical proximity" and "semantic relevance", a Gaussian kernel function is introduced to normalize and attenuate the spatial distance, and a linear weighting mechanism is used to fuse the traffic correlation coefficient. Finally, the intensity is corrected by the road function level coefficient.

[0097] II.3.1. Gaussian kernel normalization of spatial distance.

[0098] The system first introduces a Gaussian kernel function to measure physical distances based on the spatial geometric view. Perform nonlinear mapping.

[0099] This processing aims to achieve distance-normalized decay, meaning that nodes that are closer together have a higher association strength, and this association decays exponentially with increasing distance. The calculated spatial association weights... Defined as:

[0100] ;in is the bandwidth threshold of the Gaussian kernel function, with a value range of [10, 1000], used to adjust the model's sensitivity to local neighborhoods.

[0101] II.3.2. Linear weighted fusion of geographical and semantic features.

[0102] After obtaining the spatial correlation weight Correlation with traffic evolution After semantic association, the system uses a linear weighting mechanism to fuse the two views: geographic proximity and semantic relevance.

[0103] By introducing a view balance factor Its value ranges from [0, 1]. The relative importance of geometric spatial features and data evolution features in the graph composition is dynamically adjusted to obtain a preliminary comprehensive correlation strength. :

[0104] .

[0105] II.3.3. Road Functional Grade Coefficient Intensity correction.

[0106] The system further introduces a road function correction coefficient. This coefficient is used to quantify the hierarchical operational characteristics of urban road networks. It scales the connection strength of the edges based on the road's traffic level (such as expressway, arterial road, secondary arterial road, or local road).

[0107] The aim is to ensure that features of high-level road networks have higher saliency in feature extraction during subsequent graph convolution operations by adjusting the connection weights of roads at different levels in the graph structure.

[0108] In practical applications This is obtained by querying the road classification field in the electronic road network map and mapping it. For example, for expressways with a design speed greater than 80 km / h, The standard setting is 1.2; for ordinary main roads, the standard setting is 1.0.

[0109] The road function correction coefficient is designed to reflect the differences in the carrying weight of roads of different levels in congestion propagation, and its value range can be adaptively adjusted according to the road network hierarchy of a specific city.

[0110] The physical significance of setting this road function correction coefficient lies in strengthening the topological constraints of the core road network, making the simulation process of congestion propagation more consistent with the real physical evolution of traffic.

[0111] II.3.4. Final comprehensive weight formula definition.

[0112] For directed edges in the road network that have physical connections Its final comprehensive weight The calculation formula is defined as follows:

[0113] .

[0114] In the formula, For topology indicator functions, strictly limited to only Weights are calculated between physically connected node pairs; otherwise, they are set to 0, thus effectively maintaining the sparsity of the road network matrix and reducing computational complexity.

[0115] parameter The bandwidth threshold of the Gaussian kernel function is used to adjust the model's sensitivity to local neighborhood features. The smaller the value, the faster the weight decays with distance, and the model will enhance the extraction weight of features associated with nearest neighbor nodes.

[0116] The hyperparameter α takes values ​​in the range [0, 1] and serves as a view balancing factor. It dynamically adjusts the weight ratio of geometric spatial features and data evolution features in the map composition, thereby achieving the optimal feature gain balance between geographic spatial dependence and historical evolution patterns.

[0117] II.3.5. Spatiotemporal representation of road network.

[0118] Node feature vectors It is the basic data unit for spatiotemporal modeling. At each statistical time step t (e.g., a 5-minute interval), multidimensional indicators are extracted for each checkpoint node i in the road network:

[0119] Traffic flow (Flow, f): Calculated from the above steps, representing the total number of vehicles passing through the checkpoint during this period;

[0120] Average speed (V): Calculated from the above steps, it is the core indicator for determining the congestion status of a road segment.

[0121] Vectorized representation: The state of checkpoint node i at time t is represented as That is, the feature dimension F=2. This represents the total number of vehicles passing through checkpoint node i at time t. This represents the average velocity of checkpoint node i at time t.

[0122] Spatial constraints (physical constraints, such as distances, correlations, and physical topology calculated earlier) are defined using a weighted adjacency matrix W, and combined with a time series feature matrix (i.e., a three-dimensional data tensor) to construct a spatiotemporal map of road network checkpoints.

[0123] II.3.6. Training sample generation based on sliding window.

[0124] To train the diffusion model to learn the congestion propagation patterns, this invention employs a sliding window technique to segment long-sequence data (i.e., three-dimensional data tensors) into historical feature sequences. and predicting target sequences .

[0125] in Extract the continuous sequence before the current time. The observation tensors at each time step are used as input conditions for the model, and the future continuous tensors are used as input conditions. The actual observation tensor at each time step is... , as the learning objective for model generation.

[0126] Training mapping relationship: Each training sample is composed of { ,W}→ constitute.

[0127] Step 2. Construct a conditional denoising diffusion prediction model and utilize three-dimensional data tensors and The model is trained using the constructed training samples; then the trained model is used to predict future traffic evolution trajectories.

[0128] A set of pure noise data with time and spatial dimensions consistent with the prediction target is randomly selected from the standard Gaussian space.

[0129] This set of pure noise data and the latest historical feature sequences , When the number of steps and the embedding vector n are input together into the trained model, the model outputs a future traffic evolution trajectory that conforms to physical laws.

[0130] Then keep the latest historical feature sequence. and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of these represents a different trajectory in the evolution of future transportation.

[0131] Perform an inverse normalization operation on all trajectories to restore the number of vehicles and their speeds to have actual physical meaning.

[0132] like Figure 4 As shown, the conditional denoising diffusion prediction model includes a forward denoising module, a denoising network, and a reverse denoising module.

[0133] The forward noise module adds noise to the future real signal according to a preset noise scheduling sequence. Standard Gaussian noise is injected incrementally, and the arbitrary number of diffusion steps is calculated using the reparameterization formula. Noisy signals Specifically:

[0134] In the forward noise addition stage, a Markov chain is used to utilize the transition probability. Slide the window to reveal the true future signal. Gradually inject standard Gaussian noise, The temporal evolution and spatial topological features of the original traffic flow are smoothly masked within each diffusion step until it transforms into a completely disordered noise state. .

[0135] It comes from future real signals containing flow and speed characteristics. Starting from the preset number of diffusion steps Above, according to the preset noise scheduling sequence Gradually towards Gaussian noise is added to gradually mask the temporal patterns and spatial topological features of the original traffic flow, resulting in a series of noisy graph signals.

[0136] For each time step There exists a linear combination formula:

[0137] ; in The noise is standard Gaussian noise sampled independently. This future real signal is used during model training. Using the predicted target sequence from the training samples .

[0138] Firstly, according to Calculate the signal retention ratio at each step size A global decay coefficient is generated by performing a sequence accumulation operation on this ratio. (For steps from step 1 to step n) All values ​​undergo a sequence multiplication operation.

[0139] This coefficient is and The form of is used as a weighting factor, which directly affects the linear combination formula and quantitatively specifies the mixing ratio between the original traffic flow characteristics (flow and speed) and random Gaussian noise under any noise addition step number n.

[0140] The core architecture of the conditional denoising diffusion prediction model is a spatiotemporal graph symmetric denoising network ST-GUNet. This model follows the reverse iterative reasoning logic of the diffusion model on a macroscopic level, and uses spatiotemporal residual blocks as basic operators on a microscopic level.

[0141] The input to the ST-GUNet denoising network module is the current noisy signal. Historical feature sequences segmented from historical three-dimensional tensors The input includes the weighted adjacency matrix W, and the common input also includes the number of diffusion steps n; the output is a noise estimate. .

[0142] like Figure 5As shown, this invention constructs a spatiotemporal graph symmetric denoising network ST-GUNet as the core architecture. This figure illustrates in detail the microscopic operator flow of the core denoising network, which adopts an encoder-decoder symmetric architecture and fuses noisy signals through a splicing layer. It uses conditional signals and performs feature compression and multi-scale restoration using multi-layer spatiotemporal residual blocks.

[0143] like Figure 6 This is an enlarged schematic diagram of a single spatiotemporal residual block, showing its internally integrated two-layer gated causal convolution (capturing temporal dependencies), graph convolutional layer (capturing spatial topological constraints based on the W matrix), and noise progress embedding layer (Emb). The embedding vector is injected into the feature stream through an additive operator to achieve adaptive perception of different diffusion depths.

[0144] This network integrates multi-layer graph convolutional modules (GCNs) and gated recurrent units (GRUs) within a typical U-shaped structure, thereby achieving deep fusion perception of static topological constraints and dynamic temporal characteristics of the road network. Through this symmetric encoder-decoder structure, ST-GUNet can efficiently extract and reconstruct complex congestion diffusion patterns from noisy signals.

[0145] This network possesses the following characteristics: it uses a weighted adjacency matrix W as input and employs a graph convolutional neural network (GCN) to extract spatial dependencies; it uses historical feature sequences... The time step encoding n is used as a conditional input, and the time dependency is characterized by a gated recurrent unit (GRU). A symmetric encoder-decoder structure is adopted, and multi-scale spatiotemporal features are fused through skip connections during feature encoding and decoding to enhance the ability to capture patterns of different time scales and spatial ranges.

[0146] In its specific implementation, the denoising network constructed in this invention... It adopts a symmetrical encoder-decoder structure and achieves high-fidelity restoration of flow and speed indicators through multi-scale feature extraction.

[0147] Denoising Network Receive noisy signal at each diffusion step n Historical characteristic sequence The weighted adjacency matrix W and the embedding vector representing the diffusion step number n are used as joint inputs. By utilizing multi-layer feature stacking and skip connection techniques, the conditional probability distribution of future traffic signals is accurately learned under the premise of considering the road network structure, historical background and current noise level.

[0148] Through this conditional input design, the network can learn the conditional distribution of future graph signals, taking into account the road network topology, historical operating status, and current diffusion step position. The total input formula is as follows:

[0149] ;

[0150] in This is the noise estimate of the network output. For the noisy traffic map signal at the current diffusion step, Historical traffic conditions (including flow and speed characteristics). The weighted road network adjacency matrix serves as a spatial propagation constraint.

[0151] This represents the number of diffusion steps (Position Embedding). .

[0152] For the nth diffusion step, the 2ith and 2i+1th positions (forming a coordinate pair) of its position vector PE(n) are calculated as follows:

[0153] ;

[0154] ;

[0155] Where d is the total dimension of the encoding vector, and i is the index of the feature dimension. .

[0156] like Figure 5 As shown below, the processing flow of the denoising network will be described in detail.

[0157] Noisy signal , The embedding vectors of W and the diffusion step number n are used together as the global joint input.

[0158] The denoising network first utilizes the bottom splicing layer to process the noisy signal. and historical feature sequences Feature concatenation is performed, and the size of the concatenated feature sequence is (F, V, 2T), where 2T refers to the length of the concatenated sequence of the two tensors input to the denoising network.

[0159] The feature stream then enters the U-shaped architecture. In this embodiment, the U-shaped architecture includes an encoding path, a bottleneck layer, and a decoding path, with a total of five layers of spatiotemporal residual blocks, such as... Figure 5 As shown.

[0160] Each spatiotemporal residual block receives a weighted adjacency matrix W and an embedding vector of step number n simultaneously when processing the feature stream. W provides spatial topological constraints when performing graph convolution at each layer, and n enables each layer's feature stream to dynamically perceive the current denoising progress.

[0161] After being processed by the first layer of spatiotemporal residual blocks, it is transformed into a (C,V,T) tensor with hidden feature dimension C. Then it enters the second layer of spatiotemporal residual blocks, where a downsampling operation is performed in the time dimension, and the feature flow is further compressed to (C,V,T / 2).

[0162] The feature flow flows downwards into the bottleneck layer at the bottom of the U-shaped architecture, which is composed of the third layer of spatiotemporal residual blocks.

[0163] The feature stream then enters the decoding path, sequentially passing through two corresponding spatiotemporal residual blocks for temporal upsampling and restoration, and using skip connections to directly fuse the local fine-grained features of the corresponding level in the encoding stage to the corresponding level.

[0164] The high-dimensional latent features processed by the U-shaped symmetric architecture are restored to their dimensions through the top fully connected layer, outputting a dual-channel noise estimation tensor that integrates flow noise and velocity noise components.

[0165] like Figure 6 As shown, the processing flow for each spatiotemporal residual block is as follows:

[0166] Each spatiotemporal residual block first receives a feature hiding tensor H, the time series length of which changes dynamically according to the position of the block in the network; the spatiotemporal residual block then feeds the feature hiding tensor H into the first gated causal convolutional layer.

[0167] The first gated causal convolutional layer extracts dynamic dependency features of flow and velocity along the time axis, and simultaneously injects the embedding vector (Emb) transformed from the diffusion step number n into the feature stream through the addition operator to achieve adaptive perception of different denoising progress.

[0168] The fused features are fed into the second gated causal convolutional layer.

[0169] The second gated causal convolutional layer continues to extract deeper temporal dependency features along the time axis. After being processed by the first and second double-layer gated causal convolution, the features are normalized and then enter the graph convolutional layer (GCN).

[0170] The graph convolutional layer GCN synchronously accepts the weighted adjacency matrix W of the road network, uses W as a spatial topological constraint, and performs cross-node feature aggregation based on this constraint, thereby quantifying the propagation intensity of congestion on the road network path.

[0171] Finally, the original input feature hidden tensor H is fused with the spatiotemporal features output by GCN through peripheral residual connections and fed into the time scale adjustment layer. Depending on the specific needs of the network layer, the time dimension can be flexibly downsampled, upsampled, or kept unchanged to ensure that the model can capture multi-scale congestion patterns from micro fluctuations to macro trends.

[0172] The denoising network synchronously outputs a dual-channel noise estimation tensor, i.e., the predicted noise estimate, which has the same shape as the input signal. This dual-channel noise estimation tensor integrates the flow noise component and the velocity noise component.

[0173] The inverse noise reduction module converts a set of purely random noise into a single unit. As the initial state, after receiving the noise estimate output by the denoising network, the noise estimate is transformed into a denoising operator. This operator is used to perform reverse-order loop denoising calculations.

[0174] Each step removes the predicted noise components, eventually transforming the system into an ordered state that conforms to the physical laws of real traffic. In this process, different random seeds in S are input in parallel, ultimately resulting in a sample set containing S different future trajectories.

[0175] The training process of the conditional denoising diffusion prediction model in this embodiment is as follows:

[0176] I. Mathematical simulation of the forward noise addition process.

[0177] Based on a preset noise scheduling sequence, the future real signal is... (During the model training phase, i.e., predicting the target sequence) Standard Gaussian noise is gradually injected, and a noisy signal with arbitrary diffusion steps n is generated using the reparameterization formula. .

[0178] II. Conditional feature fusion at the denoising network entry point and multidimensional feature perception within the spatiotemporal residual block.

[0179] The innovative ST-GUNet module uses a concatenation operation at the input end to concatenate noisy signals. And the historical context (i.e., the sequence of historical characteristics) that serves as a guiding condition. Feature fusion is performed to map the joint signal to a high-dimensional latent feature space, generating a feature hiding tensor H for subsequent spatiotemporal feature depth perception.

[0180] H enters the spatiotemporal residual block to perform deep processing. In the time dimension, it uses gated causal convolution to capture dynamic evolution patterns and simultaneously injects step embedding vectors to perceive noise depth. At the same time, in the spatial dimension, it uses graph convolutional layers combined with a weighted matrix W to perform topological feature aggregation and quantify the propagation intensity of congestion features in complex road networks.

[0181] III. Index Co-mapping and Noise Estimation Output.

[0182] The high-dimensional latent features processed by the U-shaped symmetric architecture are then dimensionally restored through the top fully connected layer, simultaneously outputting a dual-channel noise estimation tensor with the same shape as the input signal. This tensor integrates the flow noise component and the velocity noise component. .

[0183] IV. Loss Function Construction and Parameter Optimization.

[0184] Predicting noise by calculation Compared with actual injected noise The loss function is constructed using the mean squared error (MSE) between the two values.

[0185] Meanwhile, this invention utilizes the backpropagation algorithm to update the network parameters using gradients, forcing the denoising network ST-GUNet to learn the mapping relationship that accurately identifies and removes noise features from complex spatiotemporal interference under different diffusion step sizes.

[0186] After completing model training, multi-sample prediction and future trajectory generation are performed, such as... Figure 7 As shown.

[0187] To capture the inherent randomness of traffic systems, a parallel sampling engine is driven by a single program run during the inference process, using S different random seeds to independently reconstruct S sets of distinct future evolution trajectories. .

[0188] The prediction process of the conditional denoising diffusion prediction model is as follows:

[0189] I. Initialize sampling with standard Gaussian noise.

[0190] First, regarding the full-time window to be predicted... Initial noise signals are independently and randomly sampled from standard Gaussian space. The initial noise signal has the same spatial dimension as the predicted target tensor, that is... This ensures full coverage of the features of all nodes and dimensions of the road network.

[0191] II. Iterative inverse denoising under constrained conditions.

[0192] Secondly, in the step sequence The discrete-time reverse update logic is executed; within each iteration step, the denoising network synchronously receives the current noisy signal. Historical characteristic sequence as well as As a joint constraint condition.

[0193] The denoising network perceives the road network topology constraints and the current noise progress, and accurately removes the noise residual in each iteration to drive the noisy signal to gradually move towards an ordered state that conforms to the laws of traffic physics.

[0194] III. Multi-sample parallel generation based on Monte Carlo.

[0195] To accurately capture the high degree of uncertainty in traffic flow evolution, within the same historical feature sequence And road network structure, i.e., weighted adjacency matrix Then, by changing the random seed, S independent reverse denoising samplings are performed repeatedly.

[0196] The process aims to generate a sample set containing S future traffic evolution trajectories. This allows for a complete characterization of the probability distribution of future traffic patterns within the probability space.

[0197] IV. Physical quantity reconstruction of flow rate and velocity indicators.

[0198] For each generated sample trajectory in the sample set, the traffic flow and average speed of all observed checkpoints within all future prediction steps are simultaneously included; the model output is restored to traffic parameters with actual physical meaning through inverse normalization.

[0199] A deep statistical analysis is performed on the generated multi-sample trajectory set to calculate the predicted mean and confidence interval for each node, and the operational indicators are mapped to discrete congestion levels based on preset thresholds. By statistically analyzing the frequency of different congestion levels, a quantitative assessment report is finally output, including the node congestion probability distribution, key propagation paths, and risk impact.

[0200] This invention first uses S predicted trajectories to shift from single-value predictions of traffic flow / speed to interval predictions (predicted mean and interval predictions); then it calculates the congestion level probability based on the road service level standard; next, it uses upstream and downstream checkpoints to calculate the congestion propagation probability and congestion propagation map; finally, it calculates the key nodes using a formula.

[0201] Step 3. Map the indicators of each node and each time step in the sample trajectory to discrete congestion levels, and count the frequency of occurrence of each level to obtain the probability of node congestion level and the corresponding prediction interval.

[0202] Take the set of values ​​for each checkpoint at a future preset time step, which includes... The traffic flow and driving speed obtained from the second sampling are used to calculate the arithmetic mean and variance of traffic flow and driving speed, respectively. Combined with a preset confidence level, the predicted range for traffic flow and speed fluctuations is derived. Simultaneously, a mapping transformation of state levels is performed. The driving speed and traffic flow values ​​in each trajectory are divided into discrete congestion levels, and the total number of occurrences of each congestion level in the entire sample database is counted. The total number of occurrences of each congestion level is then divided by... This yields the probability distribution of congestion levels at each checkpoint in the future.

[0203] like Figure 7 As shown, the process of calculating the probability of node congestion level is as follows:

[0204] Step 3.1. Discretization mapping of multidimensional operating indicators.

[0205] Based on the road service level classification standard, hard thresholds are applied to the average speed and traffic flow (the flow value becomes a [0-1] value after normalization) components at each time t in the multi-sample trajectory, and the operating status is divided into four discrete levels: smooth, basically smooth, lightly congested and heavily congested, thereby establishing a corresponding spatiotemporal status label matrix for each checkpoint node.

[0206] Table 1 Road Service Level Classification Standards

[0207]

[0208] Step 3.2. Frequency statistics based on the sample trajectory set.

[0209] For each checkpoint node in the road network, extract the set of traffic flow and average speed values ​​corresponding to S sample trajectories at time step t. Statistics on its future steps The frequency of falling into each congestion level;

[0210] in Indicates the first The values ​​of traffic flow and average speed corresponding to each sample trajectory.

[0211] By calculating the ratio of the number of occurrences of each level to the total number of samples S, the probability distribution prediction of the node-level congestion level is obtained.

[0212] Step 3.3. Statistical characteristics and confidence interval derivation.

[0213] For each checkpoint node in the road network, extract the set of traffic flow and average speed values ​​corresponding to S sample trajectories at time step t. The arithmetic mean is calculated by performing expectation operation on the set, and the predicted mean representing the baseline trend of future evolution is output. Simultaneously, variance operation is performed and combined with a preset confidence level (such as 95%) to calculate the prediction interval (confidence bandwidth) of each node indicator, which is used to quantify the volatility risk and uncertainty of the forecast results.

[0214] Step 4. Based on multiple future sample trajectories, statistically analyze the co-occurrence relationship of congestion states at adjacent checkpoints under different time lags, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the node congestion propagation impact degree through path probability accumulation statistics to identify key nodes of congestion propagation.

[0215] like Figure 8 As shown, the process of congestion propagation probability and key node identification is as follows:

[0216] Step 4.1. Definition and determination of spatiotemporal co-occurrence congestion events.

[0217] The linkage relationship between adjacent nodes in the road network is identified by scanning multiple sets of simulated trajectories. If multiple simulations show that "after upstream node i becomes congested, downstream node j immediately becomes congested within a very short time", then this sequence is determined to be a congestion propagation event.

[0218] Step 4.2. Statistical estimation of propagation conditional probability.

[0219] This invention uses statistical methods to quantify the intensity of congestion propagation.

[0220] Define two adjacent nodes, i above and j below. By counting the number of times that congestion occurs at all upstream nodes i, and subsequently downstream nodes j also become congested, we can calculate the single-hop conditional probability of congestion propagating from node i to node j by dividing the number of co-occurrences by the total number of times node i is congested. .

[0221] After the calculation is completed, extract all single-hop propagation conditional probabilities, calculate the arithmetic mean and standard deviation of these probabilities, add the arithmetic mean to the standard deviation by a set multiple, and set the sum as the effective propagation probability threshold.

[0222] Step 4.3. Statistics on multi-hop propagation paths and construction of risk maps.

[0223] The probability multiplication algorithm is used to track the propagation pattern of congestion over long distances and across time periods.

[0224] Multiple adjacent single-hop nodes are spliced ​​together to form a complete long-distance propagation path. The cumulative probability of congestion propagating along this path across time periods is calculated using a probability multiplication algorithm. The propagation conditional probabilities of adjacent nodes in each segment of the path are multiplied continuously, and a global risk propagation graph is constructed based on all multi-hop paths and their corresponding cumulative probabilities.

[0225] The propagation effect over long distances is quantified using a probability multiplication algorithm. For example, when congestion occurs at the source node A, the risk is calculated to propagate further to the distant node C via the transit point B, and the propagation intensity is precisely quantified by the product of the conditional probabilities of each segment.

[0226] Step 4.4. Quantitative scoring of the propagation impact of key nodes.

[0227] Based on the risk propagation map constructed in step 4.3, three physical assessment indicators are extracted.

[0228] First, the coverage of congestion triggers That is, based on the calculated effective propagation probability threshold, find all downstream nodes in the risk propagation graph whose cumulative probability exceeds the threshold, and count the total number of these effective downstream nodes as the congestion trigger coverage.

[0229] Second, the depth of the transmission path That is, the total length of the distance from node i to the farthest effective downstream node along the propagation path is measured. This length is obtained by retrieving the weighted adjacency matrix W from step 1 to obtain the actual physical distance between nodes.

[0230] Third, the cumulative risk intensity This involves extracting the traffic flow characteristics of each effective downstream node; multiplying the cumulative propagation probability of each downstream node by its traffic flow; and summing the weighted cumulative value of the probability of each propagation path and the traffic weight of the downstream node.

[0231] Specifically, the overall score of node i Calculated using the formula:

[0232] ;

[0233] in, Represents the number of nodes touched. Represents the physical length of the propagation path. This represents the cumulative value of the weighted propagation probability;

[0234] To set weighting factors, To prevent large-scale congestion. The highest value will be assigned when ensuring smooth traffic flow on the city's main roads. The highest value will be assigned when reducing overall vehicle delays. The highest value. The comprehensive score directly quantifies the potential risk of a node causing network-wide traffic operation failure; nodes with higher scores have stronger risk transmission characteristics.

[0235] Step 4.5. Integrated decision output and situation visualization.

[0236] The above quantitative scores are normalized to output a ranking of key nodes and a list of high-risk transmission paths.

[0237] The analysis results are transmitted in real time to the graphical interface of the traffic management platform. The system uses graphical representations such as probability density shaded areas or risk cloud maps to spatially display future traffic conditions.

[0238] Compared with existing methods, the present invention is significantly superior to existing solutions in at least the following aspects:

[0239] 1. The prediction method evolved from deterministic point prediction to probability distribution prediction, thereby increasing the amount of feature information.

[0240] Existing GCN-GRU-like methods typically output only a single predicted value. This invention constructs a conditional denoising diffusion model on the spatiotemporal map of checkpoints, directly learning the conditional distribution of future multi-step traffic states. While providing the predicted mean, the model can also obtain the prediction interval and congestion level probability distribution for each checkpoint at each time step through multiple samplings, evolving from deterministic point prediction to probabilistic distribution prediction, thus increasing the amount of feature information.

[0241] 2. The spatiotemporal correlation modeling is refined, which enhances the adaptability to complex road network topologies.

[0242] Traditional methods often employ fixed adjacency matrices, resulting in simplistic spatial weight designs. Furthermore, they frequently utilize rolling prediction methods in the time dimension, which can easily lead to error accumulation. This invention introduces a weighted checkpoint graph representation at the road network level, integrating road topology, spatial distance, and operational relevance to construct edge weights, making the graph structure more closely resemble realistic traffic interactions. At the model level, a conditional diffusion mechanism is used to model the entire future trajectory across multiple time steps, avoiding the error accumulation inherent in traditional rolling prediction.

[0243] 3. Directly outputting the probability of node-level congestion levels improves the interpretability of the project and the efficiency of its application.

[0244] Existing prediction models mostly output continuous flow or speed values, which still require manual threshold setting for state mapping in practical applications, resulting in a lack of consistency. This invention, based on multiple sampling trajectories, uses an embedded frequency statistics mechanism to directly map the operational indicators of each checkpoint and each time step to discrete levels, and statistically analyzes the frequency of occurrence of each level to obtain the probability of node-level congestion levels and corresponding prediction intervals. This output format eliminates the need for secondary manual processing, significantly improving the intuitiveness and usability of the prediction results.

[0245] 4. Implement congestion propagation analysis and critical node assessment within a unified framework.

[0246] Traditional congestion propagation studies rely on independent, simplified mechanistic models, such as epidemic-like propagation models, leading to a decoupling between prediction results and propagation analysis. This invention utilizes a large number of future sample trajectories generated by a diffusion model to directly perform congestion propagation analysis based on data. By statistically analyzing the co-occurrence of congestion at adjacent nodes under different time lags, the conditional probability of congestion propagation between nodes is estimated; and by accumulating path probabilities, a key node influence index is defined. Therefore, prediction, congestion probability, propagation paths, and key node identification are all completed within the same model framework, eliminating the need for separate propagation equations and ensuring the consistency of system logic.

[0247] 5. More robust to abnormal and uncertain situations, with higher stability.

[0248] Traditional deterministic models are prone to significant biases when dealing with anomalies such as non-Gaussian and heavy-tailed distributions in urban traffic systems, especially when handling sudden changes. This invention employs a diffusion model, utilizing a mechanism of forward noise addition and reverse denoising to effectively fit complex non-Gaussian distributions. During training, the model learns the ability to recover reasonable traffic conditions from high-noise environments. Even when facing abnormal fluctuations, the system can still output stable probabilistic predictions and quantify uncertainty using prediction intervals. The generative probabilistic model demonstrates superior adaptability to data distribution patterns compared to conventional mean squared error regression models.

[0249] Example 2

[0250] This embodiment 2 describes a system for predicting urban road network traffic conditions and congestion propagation probability. This system has the same inventive concept as the method for predicting urban road network traffic conditions and congestion propagation probability in embodiment 1 above.

[0251] The urban road network traffic status and congestion propagation probability prediction system in this embodiment includes the following modules:

[0252] The preprocessing module is used to acquire checkpoint data. Through data preprocessing, it obtains complete and continuous checkpoint operation time series data, and obtains a three-dimensional data tensor through time aggregation and index calculation; at the same time, it constructs a weighted adjacency matrix. ;

[0253] The trajectory prediction module is used to construct a conditional denoising diffusion prediction model, utilizing three-dimensional data tensors and... The model is trained using the constructed training samples; then the trained model is used to predict future traffic evolution trajectories.

[0254] First, a set of pure noise data with the same time and spatial dimensions as the prediction target is randomly selected from the standard Gaussian space;

[0255] This set of pure noise data and the latest historical feature sequences , When the number of steps and the embedding vector n are input together into the trained model, the model outputs a future traffic evolution trajectory that conforms to physical laws.

[0256] Then keep the latest historical feature sequence. and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of the following is a unique future transportation evolution trajectory, i.e., a sample trajectory.

[0257] The urban road network traffic condition prediction module is used to map the index of each node and each time step in the sample trajectory to discrete congestion levels, and to count the frequency of occurrence of each level, thereby obtaining the probability of node congestion level and the corresponding prediction interval.

[0258] The module also includes a congestion propagation probability prediction module, which is used to statistically analyze the co-occurrence relationship of congestion states at different time lags of adjacent checkpoints based on multiple future sample trajectories, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the impact degree of node congestion propagation through path probability accumulation statistics, and identify key nodes of congestion propagation.

[0259] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

[0260] Example 3

[0261] This embodiment 3 describes a computer device, which includes a memory and one or more processors. Executable code is stored in the memory, and when the processor executes the executable code, it implements steps for a method to predict urban road network traffic conditions and congestion propagation probability.

[0262] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.

[0263] Example 4

[0264] This embodiment 4 describes a computer-readable storage medium storing a program that, when executed by a processor, implements steps for a method to predict urban road network traffic conditions and congestion propagation probability.

[0265] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.

[0266] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.

Claims

1. A method for predicting urban road network traffic conditions and congestion propagation probability, characterized in that, Includes the following steps: Step 1. Obtain checkpoint data. Preprocess the data to obtain complete and continuous checkpoint operation time series data. Obtain a three-dimensional data tensor through time aggregation and indicator calculation. Simultaneously, construct a weighted adjacency matrix. ; Step 2. Construct a conditional denoising diffusion prediction model and utilize three-dimensional data tensors and The training samples are constructed, the model is trained, and then the trained model is used to predict future traffic evolution trajectories. First, a set of pure noise data with the same time and spatial dimensions as the prediction target is randomly selected from the standard Gaussian space; Then, this set of pure noise data and the latest historical feature sequence... Weighted adjacency matrix The step number embedding vector n is input into the trained model, and the model outputs a future traffic evolution trajectory that conforms to physical laws. Maintain the latest historical feature sequence and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of the following is a unique future transportation evolution trajectory, i.e., a sample trajectory. Step 3. Map the index of each node and each time step in the sample trajectory to a discrete congestion level, and count the frequency of occurrence of each level to obtain the probability of node congestion level and the corresponding prediction interval. Step 4. Based on multiple future sample trajectories, statistically analyze the co-occurrence relationship of congestion states at adjacent checkpoints under different time lags, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the impact degree of node congestion propagation through path probability accumulation statistics to identify key nodes of congestion propagation.

2. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 1, characterized in that, In step 1, the three-dimensional data tensor is represented as: ; Among them, the first dimension This represents a historical time series divided according to fixed time intervals, i.e., preset step sizes; the second dimension... Represents the set of all observation checkpoints within the region; third dimension It represents two core operational indicators: traffic flow and average speed. Traffic flow within each time window refers to the total number of vehicles passing through that checkpoint within that time window; average speed within each time window refers to the average speed of all vehicles passing through two adjacent checkpoints within that time window.

3. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 1, characterized in that, In step 1, the weighted adjacency matrix The construction process is as follows: I. Establish the initial topology; Each checkpoint is considered a node in a graph, and all checkpoints constitute a set of nodes. ; in Total number of checkpoints; , … Each node is represented separately. Based on urban road centerline data and traffic flow information, if vehicles exit from the checkpoint... After exiting, it can reach the checkpoint directly without going through other checkpoints. Then it is determined that there exists a directed edge. ; A basic topological adjacency matrix is ​​constructed based on the above determination rules. ; in , indicating checkpoint With checkpoint There is a connection. This indicates that there is no direct connection; II. Calculation of overall weights; In the basic topological adjacency matrix Based on this, spatial distance is introduced. Traffic correlation and road function correction coefficient Construct a weighted adjacency matrix ,in medium elements The formula is expressed as follows: ； Where α is a hyperparameter, and its value ranges from [0, 1]. The bandwidth threshold for the Gaussian kernel function; For topology indicator functions, only Weights are calculated between nodes that are physically connected; otherwise, they are set to 0.

4. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 1, characterized in that, The conditional denoising diffusion prediction model includes a forward denoising module, a denoising network, and a reverse denoising module. The forward noise module adds noise to the future real signal according to a preset noise scheduling sequence. Standard Gaussian noise is injected incrementally, and the arbitrary number of diffusion steps is calculated using the reparameterization formula. Noisy signals ; The denoising network adopts a symmetrical encoder-decoder structure; The denoising network at each diffusion step Receive noisy signals Historical characteristic sequence Weighted adjacency matrix The embedding vector representing the diffusion step number n is used as joint input. By utilizing multi-layer feature stacking and skip connection techniques, the conditional probability distribution of future traffic signals is learned under the premise of considering the road network structure, historical background and current noise level. The denoising network outputs a dual-channel noise estimation tensor, which is consistent with the shape of the input signal, and is the predicted noise estimate. This dual-channel noise estimation tensor integrates the flow noise component and the velocity noise component. The reverse denoising module takes a set of pure random noise as the starting state, receives the noise estimate output by the denoising network, converts the noise estimate into a denoising operator, and uses the denoising operator to perform reverse cyclic denoising calculation. Each step removes the predicted noise components, eventually transforming the system into an ordered state that conforms to the physical laws of real traffic. In this process, different random seeds in S are input in parallel, ultimately resulting in a sample set containing S different future trajectories.

5. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 4, characterized in that, The denoising network adopts a symmetrical U-shaped encoder-decoder architecture, and its processing flow is as follows: The denoising network first utilizes the bottom splicing layer to process the noisy signal. and historical feature sequences Feature concatenation is performed, and the size of the concatenated feature sequence is (F, V, 2T), where 2T refers to the length of the concatenated sequence of the two tensors input to the denoising network. The concatenated feature stream then enters the U-shaped architecture; the U-shaped architecture consists of three parts: the encoding path, the bottleneck layer, and the decoding path, with a total of five spatiotemporal residual blocks; each spatiotemporal residual block receives the weighted adjacency matrix W and the embedding vector of the diffusion step number n simultaneously when processing the feature stream, where W is used to provide spatial topological constraints when performing graph convolution at each layer, and n is used to enable each layer of feature stream to dynamically perceive the current denoising progress; After being processed by the first layer of spatiotemporal residual blocks, it is transformed into a (C,V,T) tensor with hidden feature dimension C. Then it enters the second layer of spatiotemporal residual blocks, where a downsampling operation is performed in the time dimension, and the feature flow is further compressed to (C,V,T / 2). The feature flow flows downwards into the bottleneck layer at the bottom of the U-shaped architecture, which is composed of the third layer of spatiotemporal residual blocks; The feature stream then enters the decoding path, passes through two corresponding spatiotemporal residual blocks in sequence for temporal upsampling and restoration, and uses skip connections to directly fuse the local fine-grained features of the corresponding level in the encoding stage to the corresponding level. The high-dimensional latent features processed by the U-shaped symmetric architecture are restored to their dimensions through the top fully connected layer, outputting a dual-channel noise estimation tensor that integrates flow noise and velocity noise components.

6. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 5, characterized in that, The processing flow for the spatiotemporal residual block is as follows: Each spatiotemporal residual block first receives the feature hiding tensor H, and then feeds H into the first gated causal convolutional layer; The first gated causal convolutional layer extracts the dynamic dependency features of flow and velocity along the time axis, and simultaneously injects the embedding vector transformed from the diffusion step number n into the feature stream through the addition operator to achieve adaptive perception of different denoising progress. Next, the fused features are fed into the second gated causal convolutional layer; The second gated causal convolutional layer continues to extract deep temporal dependency features along the time axis; After being processed by two layers of gated causal convolution, the features are normalized and then enter the graph convolutional layer GCN. The graph convolutional layer GCN synchronously accepts the road network weighted adjacency matrix W, uses W as a spatial topological constraint, and performs cross-node feature aggregation based on this constraint, thereby quantifying the propagation intensity of congestion on the road network path. Finally, the original input feature hidden tensor H is fused with the spatiotemporal features output by GCN through peripheral residual connections and fed into the time scale adjustment layer to perform downsampling, upsampling or maintain the original size in the time dimension.

7. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 4, characterized in that, The training process of the conditional denoising diffusion prediction model is as follows: Before model training, the long sequence data, i.e., the three-dimensional data tensor obtained in step 1, is first segmented using the sliding window technique into historical feature sequences. and predicting target sequences ; in Extract the continuous sequence before the current time. The observation tensors at each time step are used as input conditions for the model, and the future continuous tensors are used as input conditions. The actual observation tensor at each time step is... , as the learning objective for model generation; The forward diffusion process of the model will propagate the data to the future real data, i.e., the predicted target sequence. Gaussian noise is gradually injected into the system. During training, a number is randomly selected as the current diffusion step number n, generating a noisy signal at the current diffusion step number. ; Noise data obtained from the model's forward process Extracted historical feature sequences The weighted adjacency matrix W and the step embedding vector n representing the current diffusion progress are input together into the conditional denoising diffusion model for training. After integrating these four pieces of information internally, the model outputs a noise prediction value, namely the dual-channel noise estimation tensor. The loss function is constructed by comparing the mean square error between the prediction value and the actual injected noise. The backpropagation algorithm is used to update the network parameters of the conditional denoising diffusion prediction network by gradient, which forces the denoising network to learn the mapping relationship of accurately identifying and stripping noise features from complex spatiotemporal interference under different diffusion steps.

8. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 4, characterized in that, The prediction process of the conditional denoising diffusion prediction model is as follows: First, regarding the full-time window to be predicted... Initial noise signals are independently and randomly sampled from standard Gaussian space. The spatial dimension of the initial noise signal is completely consistent with the tensor of the predicted target, that is... ; Secondly, in the step sequence The discrete-time reverse update logic is executed; within each iteration step, the denoising network synchronously receives the current noisy signal. Historical characteristic sequence as well as As a joint constraint; The denoising network perceives the road network topology constraints and the current noise progress, and accurately removes the noise residual in each iteration to drive the noisy signal to gradually move towards an ordered state that conforms to the laws of traffic physics. To accurately capture the high degree of uncertainty in traffic flow evolution, within the same historical feature sequence And road network structure, i.e., weighted adjacency matrix Next, by changing the random seed, S independent reverse denoising samplings are performed repeatedly; The process aims to generate a sample set containing S future traffic evolution trajectories. ; For each generated sample trajectory in the sample set, the traffic flow and average speed of all observed checkpoints within all future prediction steps are simultaneously included; the model output is restored to traffic parameters with actual physical meaning through inverse normalization.

9. The method for predicting urban road network traffic conditions and congestion propagation probability according to claim 1, characterized in that, In step 3, the process of calculating the node congestion level probability is as follows: Step 3.

1. Discretization mapping of multidimensional operating indicators; Based on the road service level classification standard, hard thresholds are applied to the average speed and traffic flow components at each time t in the multi-sample trajectory to divide the operating status into four discrete levels: smooth, basically smooth, lightly congested, and heavily congested. Step 3.

2. Frequency statistics based on the sample trajectory set; For each checkpoint node in the road network, extract the set of traffic flow and average speed values ​​corresponding to S sample trajectories at time step t. Statistics on its future steps The frequency of falling into each congestion level; in Indicates the first The traffic flow and average speed values ​​corresponding to each sample trajectory; By calculating the ratio of the number of occurrences of each level to the total number of samples S, the probability distribution prediction of the node-level congestion level is obtained. Step 3.

3. Statistical characteristics and confidence interval derivation; For each checkpoint node in the road network, extract the set of traffic flow and average speed values ​​corresponding to S sample trajectories at time step t. The arithmetic mean is calculated by performing the expected operation on the set; Output the predicted mean representing the baseline trend of future evolution; simultaneously perform variance calculation and combine it with the preset confidence level to calculate the prediction interval of each node indicator, so as to quantify the volatility risk and uncertainty of the forecast results; In step 4, the process of congestion propagation probability and key node identification is as follows: Step 4.

1. Definition and determination of spatiotemporal co-occurrence congestion events; By scanning multiple sets of simulated trajectories, the linkage relationship between adjacent nodes in the road network is identified. The continuous process in which upstream node i becomes congested and downstream node j subsequently becomes congested is determined as a spatiotemporal co-occurrence congestion event. Step 4.

2. Statistical estimation of propagation conditional probability; Define two adjacent nodes, i above and j below. By counting the number of times that congestion occurs at all upstream nodes i, and subsequently downstream nodes j also become congested, we can calculate the single-hop conditional probability of congestion propagating from node i to node j by dividing the number of co-occurrences by the total number of times node i is congested. ; After the calculation is completed, extract all single-hop propagation conditional probabilities, calculate the arithmetic mean and standard deviation of these probabilities, add the arithmetic mean to the standard deviation by a set multiple, and set the sum as the effective propagation probability threshold. Step 4.

3. Multi-hop propagation path statistics and risk map construction; Multiple adjacent single-hop nodes are spliced ​​together to form a complete long-distance propagation path. The cumulative probability of congestion propagating along the path across time periods is calculated using a probability multiplication algorithm. The propagation conditional probabilities of adjacent nodes in each segment of the path are multiplied continuously, and a global risk propagation graph is constructed based on all multi-hop paths and their corresponding cumulative probabilities. Step 4.

4. Quantitative scoring of the propagation impact of key nodes; Based on the risk propagation map constructed in step 4.3, three physical assessment indicators are extracted; First, the coverage of congestion triggers That is, based on the calculated effective propagation probability threshold, find all downstream nodes whose cumulative probability exceeds the threshold from the risk propagation map, and count the total number of these effective downstream nodes as the congestion trigger coverage. Second, the depth of the transmission path That is, the total length of the distance from node i to the farthest effective downstream node along the propagation path is measured. This length is obtained by retrieving the weighted adjacency matrix W from step 1 to obtain the actual physical distance between nodes. Third, the cumulative risk intensity That is, extract the traffic flow characteristics of each effective downstream node; multiply the cumulative propagation probability of each downstream node by its traffic flow and sum the weighted cumulative value of the probability of each propagation path and the traffic weight of the downstream node. Specifically, the overall score of node i Calculated using the formula: ;in Represents the number of nodes touched. Represents the physical length of the propagation path. This represents the cumulative value of the weighted propagation probability; To set weighting factors, ; Step 4.

5. Normalize the above quantitative scores and output the ranking of key nodes and the list of high-risk transmission paths.

10. A system for predicting urban road network traffic conditions and congestion propagation probability, characterized in that, Includes the following modules: The preprocessing module is used to acquire checkpoint data. Through data preprocessing, it obtains complete and continuous checkpoint operation time series data, and obtains a three-dimensional data tensor through time aggregation and index calculation; at the same time, it constructs a weighted adjacency matrix. ; The trajectory prediction module is used to construct a conditional denoising diffusion prediction model, utilizing three-dimensional data tensors and... The model is trained using the constructed training samples; then the trained model is used to predict future traffic evolution trajectories. First, a set of pure noise data with the same time and spatial dimensions as the prediction target is randomly selected from the standard Gaussian space; This set of pure noise data and the latest historical feature sequences , When the number of steps and the embedding vector n are input together into the trained model, the model outputs a future traffic evolution trajectory that conforms to physical laws. Then keep the latest historical feature sequence. and By changing the random seed used during initialization, the model can be run repeatedly without changing the initial seed. The noise reduction process is repeated to obtain Each of the following is a unique future transportation evolution trajectory, i.e., a sample trajectory. The urban road network traffic condition prediction module is used to map the index of each node and each time step in the sample trajectory to discrete congestion levels, and to count the frequency of occurrence of each level, thereby obtaining the probability of node congestion level and the corresponding prediction interval. The module also includes a congestion propagation probability prediction module, which is used to statistically analyze the co-occurrence relationship of congestion states at different time lags of adjacent checkpoints based on multiple future sample trajectories, estimate the conditional probability of congestion propagating from upstream nodes to downstream nodes, and define the impact degree of node congestion propagation through path probability accumulation statistics, and identify key nodes of congestion propagation.

Images